How We Built It

Our React SPA served the same empty HTML shell for every URL. Google couldn't see unique titles, descriptions, or content - just JavaScript that needed to run first.

Astro pre-renders thousands of static HTML pages with full SEO meta tags. React components only hydrate where interactivity is needed, such as tabs, quizzes, and paper viewers.

Each generated page has unique titles, descriptions, Open Graph tags, Twitter cards, and canonical URLs. Secondary and O-Level pages now get the same SEO treatment as Primary pages.

A signed, HttpOnly cookie remembers that a visitor has completed the access prompt. It does not contain the visitor's email address.

A small Node API validates the cookie using a separate PostgreSQL database. Registration records store name, email, consent version, signup time, and last access time; linked resource records aggregate the levels, subjects, quizzes, and papers viewed.

Canonical resource pages include questions and answers in their initial HTML, alongside title, description, canonical URL, breadcrumbs, and structured data identifying the registration-gated section. Raw generated Markdown URLs remain blocked.

Visitors can remove their matching registration and linked resource history through the unsubscribe page. The Privacy Policy explains the collection purpose and contact route.

Recurring issues below 8.0 now feed the Stage 2-1, 2-2, 3, 4, and 5 prompts: exact question counts, complete passages and data, exact mark totals, level calibration, full answer keys, and cleaner LaTeX/notation.

Visual-dependent questions may keep image work for Stage 3-3 or Stage 5-3, but they must now use explicit <image_placeholder>...</image_placeholder> tags with labels, values, linked question number, and required visual features. Stage 5-2 image manifest converts those tags into a CSV job table before images are generated or injected.

Stage 9 exposed older placeholder-only papers and weak source/data questions. The newer prompts now ban vague placeholders and require each quiz or paper to be answerable from the student-facing markdown once tagged images are generated.

Stage 9-2 tracks stage, model, date, subject, level, syllabus year, benchmark score, local markdown path, answer path, missing-image signals, and whether the next action is a new generation, repair, benchmark, or image follow-up.

Each level gets exactly what they need. Math syllabus covers P1-P6, but we split it into 6 separate files for easier consumption.

Science is not taught in P1-P2. Our pipeline respects this - no science content for those levels.

Some secondary PDFs bundle many schools into one file. We keep the original OCR intact, then create per-school files beside the original bundle markdown: 83 bundles normalized into 844 per-school/front-matter markdown files. Page validation matched exactly: 29,120 source pages and 29,120 split pages.

Some compiled PDFs place marking schemes much later than the question paper. The split groups delayed answer keys back into the same school file instead of blindly cutting only contiguous ranges.

Used for primary content and earlier quiz/paper generation where quality mattered more than cost.

From 1 June 2026, Secondary, O-Level, and A-Level quizzes and papers are generated in parallel with these models to reduce cost and preserve model provenance on the website. In practice, different LLMs run at different speeds, and their effective handling of long syllabus and exam-paper context can differ even when the prompt contract is the same.

Fast, cost-effective for bulk processing. Good at recognizing question structure.

The first OCR-to-markdown generation used both Gemini and Pixtral. Pixtral was useful for formula-heavy and table-heavy syllabus pages; Gemini 2.0 was better for school exam-paper OCR, especially scanned PDFs and non-standard page imaging.

Geometry proofs, physics problems, multi-step calculations. When you need to show your working.

Singapore English (en-SG) and Mandarin Chinese (zh-CN) voices. Custom pitch and rate settings per character for natural-sounding dialogue.